Surprising Absence of Co-Occurring Gene Alterations in A Recent Pan-Cancer Study

(First published on LinkedIn, 2/2017)

The genetic theory of cancer postulates that distinct, sequential DNA alterations over time promote oncogenesis and tumor evolution.

'Genetic' is used here as a general term to describe the many changes found in the DNA of sporadic tumors.

The DNA alterations in cancer are numerous. Many of these changes consist of point mutations, gene copy number gains or losses, deletions, insertions, and translocations. Researchers routinely detect the co-occurrence of multiple somatic mutations in tumors. Given the genomic complexity and chromosomal instability found in many cancers, the prevailing idea is that single gene hits are not sufficient to explain oncogenesis and tumor evolution. Other changes, described as epigenetic, are likely as important but typically not detected using standard DNA-sequencing methods only.

A large body of preclinical and clinical data support the genetic theory. Targeting 'druggable' proteins derived from altered genes has improved clinical outcomes when compared with traditional chemotherapy. Much of the excitement around 'precision medicine' is based on drugs developed for patients with mutations, amplifications, and translocations. Unfortunately, most patients who respond initially to targeted drugs become refractory over time. Clinicians have also started to explore the use of drug combinations targeting multiple gene alterations/pathways. The hope is that this will improve efficacy and, potentially, overcome drug-resistance.

If oncogenic addiction is the result of some Darwinian selection within the organism, it should be possible to identify very specific co-occurring 'genotypes' in large genomic tumor studies. Various studies have supported this view by reporting thousands of co-occurrences using standard statistical tests. Other approaches seek out biologically relevant genes by testing for mutually exclusive gene alterations. Exclusivity is the opposite of co-occurrence and suggests that 'exclusive' gene alterations potentially exert significant 'stand-alone' oncogenic activity and, therefore, do not require co-occurrence (=synergy). Redundancy in some pathways could also explain the onset of resistance in some tumors by activating escape pathways.

A recent publication in Genome Biology suggests that many of the reported co-occurrences may be spurious and could be explained by chance alone when tested more rigorously. In addition, it suggests that some of the frequently used tests are too stringent to allow for the efficient discovery of mutually exclusive genes.

It would be helpful, of course, to separate co-occurrences from statistical noise. We could then identify novel combinations of oncogenic 'actors' more objectively and define innovative patient selection strategies. Such efforts, should, in principle, also confirm many of the 100+ known oncogenes and aid the discovery of currently unknown driver mutations.

The authors' starting point (Canisius et al. Genome Biology (2016) 17:261) is the observation that statisticians who test for co-occurrence assume an unrealistic independent and identical distribution (i.i.d.) of the gene alterations across the different tumor types. The authors then demonstrate, via simulation, that the i.i.d. assumption inherently results in high false-positive rates (FDR). They then propose a new algorithm called DISCOVER (Discrete Independence Statistic Controlling for Observations with Varying Event Rates). DISCOVER does away with the i.i.d assumption and instead incorporates tumor-specific event rates for testing co-occurrence and mutual exclusivity. Interestingly, DISCOVER accurately identifies various altered genes known to have 'natural' physical association or linkage because of their proximity in amplified DNA fragments or their co-location in chromosomal arms (a natural internal control for testing the algorithm). However, after applying DISCOVER to 3386 tumors studied in the ATCG pan-cancer initiative representing a total of 12 different cancers types, the result is rather surprising:

Although DISCOVER identifies a list of 100+ mutually-exclusive genes, many of which map to known canonical pathways, it found only one significant co-occurrence between TP53 mutations and MYC amplification after increasing the FDR rate from 1% to 3%.

What are the possible explanations? The obvious one is that there no substantial statistical evidence for co-occurrence. Previously reported associations were numerous, but these findings may need to be reviewed if the underlying statistical assumptions were flawed (i.i.d.). The study, on the other hand, only evaluates copy number gain/loss and mutations as the major alterations. Naturally, epigenetic changes cannot be the subject of this study. Nevertheless, the absence of statistically significant co-occurrence of what are considered widespread tumorigenic alterations mechanisms is very surprising.

Looking for mutually exclusive genes, DISCOVER detects many of the known oncogenes. KRAS mutations appear mutually exclusive with ESR1, TP53, BRAF and EGFR. The authors find that PTEN mutations are mutually exclusive with ERBB2 gain, CDKN2B loss, CDKN2A loss, and PI3KCA mutations. KRAS and NRAS also appear mutually exclusive. MYC mutations are found to be mutually exclusive with PI3CA mutations perhaps suggesting the need to evaluate MYC status in patients selected for PI3CA inhibitor treatment. Similarly, MYC gain and MDM4 gain are mutually exclusive as are MYC and CCND1 gains. A detailed list of the 100+ mutually exclusive genes can be found in the addendum to the paper (see reference).

Traditionally, we design patient selection and treatment strategies for new drugs often using small sets (n-size) of pre-clinical and clinical data. Pan-cancer genomic surveys such as the one presented in this paper may lead to a more objective understanding of the association between altered genes and could, therefore, help to base clinical tumor genetics, drug development, and patient selection strategies on a firmer statistical footing. While DISCOVER is clearly able to identify mutually exclusive genes that map to various canonical pathways, the absence of co-occurrences above what the authors consider statistical noise (chance) is troubling and may require a rethinking of the importance of co-occurrence in the context of tumor genesis and evolution.

Journal reference and images: Canisius et al. Genome Biology (2016) 17:261